Graph-based Method for Detecting Fraud Regarding Activity Relationships in Business Process Models

Process-Based Fraud (PBF) poses a threat to organizational integrity, resulting from deviations in the execution of business processes compared to established standard operating procedures (SOPs). Unlike transactional fraud, which targets financial discrepancies, PBF focuses on control-flow inconsistencies that may lead to unauthorized actions, policy violations, or operational inefficiencies. This study proposes a novel graph-based architecture for PBF detection using Neo4j, integrating three distinct graph-based methods: two similarity-based methods using Jaccard similarity and Levenshtein distance, and one logic-based method using PBF-type-specific rules. Embedded within a graph-native environment, this approach enables the detection of several underexplored PBF types, including skipped activities, wrong patterns, and wrong decisions. An experimental evaluation of 1,000 cases reveals that the graph-based method with Levenshtein distance achieves the best performance in detecting incorrect patterns, achieving perfect performance with an F1-score of 1.00. For wrong decision detection, the graph-based method with rule-based logic attained an F1-score of 1.00. In identifying skipped activities, both graph-based methods, using Jaccard and Levenshtein, achieved F1-scores of 0.98 at a similarity threshold of 1.0. These outcomes demonstrate that the proposed graph-based approach provides an effective solution for accurately detecting various types of process-based fraud in business process models.